Loading...

Media is loading
 

Start Date

30-4-2021 12:00 AM

Document Type

Oral Presentation

Department

Engineering

Faculty Mentor

Asheesh Lanba, PhD

Keywords

LATscan, Artificial Intelligenc, e Neural Network, UNET, RootPainter, Ilastik, Soybean Segmentation, Computer Vision

Abstract

Improvements in contemporary machine learning architectures have drastically increased the available tools in biological analysis, particularly in feature segmentation in 2D images and 3D models. A major obstacle for building machine learning architectures is associated with generating large amounts of training data, a process which requires significant time and processing resources. Recent software tools allow for an interactive training approach that show great promise in reducing the time and effort needed to generate training data. This study analyzes cross-sectional images of Soybean stems obtained via Laser Ablation Tomography (LATscan) using two open-source interactive machine-learning tools. Both software tools are trained to identify Xylem conduits depicted in the LATscan images, from which count and area data are collected. This resulting information is then compared to identify any possible bias in the algorithms and training methods. The results show effective creation of predictive models with a low amount of training data, thus presenting plant scientists with a powerful new tool for quantifying plant features that can be used for high impact phenotyping studies.

TM2021_Elias_transcript.txt (6 kB)
Interactive Machine Learning Methods for the Quantification of Vascular Features in Soybean Images Obtained via Laser Ablation Tomography (LATscan) - transcript

Open Access?

1

Share

COinS
 
Apr 30th, 12:00 AM

Interactive Machine Learning Methods for the Quantification of Vascular Features in Soybean Images Obtained via Laser Ablation Tomography (LATscan)

Improvements in contemporary machine learning architectures have drastically increased the available tools in biological analysis, particularly in feature segmentation in 2D images and 3D models. A major obstacle for building machine learning architectures is associated with generating large amounts of training data, a process which requires significant time and processing resources. Recent software tools allow for an interactive training approach that show great promise in reducing the time and effort needed to generate training data. This study analyzes cross-sectional images of Soybean stems obtained via Laser Ablation Tomography (LATscan) using two open-source interactive machine-learning tools. Both software tools are trained to identify Xylem conduits depicted in the LATscan images, from which count and area data are collected. This resulting information is then compared to identify any possible bias in the algorithms and training methods. The results show effective creation of predictive models with a low amount of training data, thus presenting plant scientists with a powerful new tool for quantifying plant features that can be used for high impact phenotyping studies.

blog comments powered by Disqus